|Publication number||US4897811 A|
|Application number||US 07/145,630|
|Publication date||Jan 30, 1990|
|Filing date||Jan 19, 1988|
|Priority date||Jan 19, 1988|
|Also published as||DE68924078D1, EP0325119A2, EP0325119A3, EP0325119B1|
|Publication number||07145630, 145630, US 4897811 A, US 4897811A, US-A-4897811, US4897811 A, US4897811A|
|Inventors||Christopher L. Scofield|
|Original Assignee||Nestor, Inc.|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (2), Non-Patent Citations (2), Referenced by (92), Classifications (14), Legal Events (7)|
|External Links: USPTO, USPTO Assignment, Espacenet|
δωnm =(+/-)|x(t)-x(t+1)|-(L+2)Δnm (f(t),f(t+1))
General learning algorithms for multi-layer neural networks have only recently been studied in detail. Recent work in this area began with the multi-layer Boltzmann machines. Hinton, G. E., and Sejnowski: "Optimal Perceptual Inference," Proceedings of the IEEE Conference on Computer Vision Pattern Recognition, 448-453 (1983), and Hinton, G. E., Sejnowski, T. J., Ackley, D. H.: "Boltzmann Machines: Constraint Satifaction Networks that Learn", Tech. Report CMU-CS-84-119, Carnegie-Mellon University (May 1984). Hinton and Sejnowski found that a learning algorithm for multi-layer networks could be described for their system. However, the low rate of convergence of the synaptic state of this system led them, and others to look for alternative multi-layer learning systems. Rumelhart, Hinton and Williams found that a generalization of the delta rule could describe learning in multi-layer feedforward networks. Rumelhart, D. E., Hinton, G. E., and Williams, R. J.: "Learning Representations by Back Propagating Errors", Nature, 323, 533-536 (1986). This delta rule was independently developed by Parker. Parker, D. B.: "Learning-logic (TR-47)," Massachusetts Institute of Technology, Center for Computational Research in Economics and Management Science, (1985). This system, now often called "Back Propagation", is much faster than the Boltzmann machine and is able to automatically acquire internal synaptic states which seem to solve many of the classic toy problems first posed by Minsky and Papert. Minsky, M. and Papert, S.: Perceptrons, MIT Press (1969). These complex internal states have been called "internal representations" of the pattern environment. Rumelhart, D. E., Hinton, G. E., and Williams, R. J.: "Learning Internal Representations by Error Propagation," in D. E. Rumelhart and J. L. McClelland (Eds.) Parallel Distributed Processing, MIT Press, 318-364 (1986). However, it has been found that the convergence rate of the synaptic state of this system goes much slower than linearly with the number of layers in the network. Ballard, D. H.: "Modular Learning in Neural Networks," Proceedings of the Sixth National Conference on Artificial Intelligence, 1, 279-284 (1987). This property has been called the "scaling problem" since it appears to be a significant limitation on the scaling of such networks to large, real-world, problems. Hinton, G. E., and Sejnowski: "Neural Network Architectures for AI," Sixth National Conference on Artificial Intelligence, Tutorial Program MP-2(1987). In the aforementioned article on "Modular Learning", Ballard proposed a method for handling the scaling problem by stacking auto-associating units one on the other. This method violates the feedforward architecture, but the system does appear to reduce the multi-layer learning time.
The Boltzmann machine was an extension of the work of Hopfield who had been studying single layer recurrent networks. Hopfield, J. J.: "Neural Networks and Physical Systems with Emergent Collective Computational Abilities," Proc. Natl. Acad. Sci. U.S.A. 79, 2554-2558 (April 1982), and Hopfield, J. J.: "Neurons with Graded Response Have Collective Computational Properties Like Those of Two-State Neurons," Proc. Natl. Acad. Sci. U.S.A. 81, 2088-3092 (May 1984). Hopfield introduced a method for the analysis of settling of activity in recurrent networks. This method defined the network as a dynamical system for which a global function called the "energy" (really a Liapunov function for the autonomous system describing the state of the network) could be defined. This energy then contained fixed points in the system state space. Hopfield showed that flow in state space is always toward the fixed points of the dynamical system if the matrix of recurrent connections satisfies certain conditions. With this property, Hopfield was able to define the fixed points as the sites of memories of network activity.
Like its forerunners, the Hopfield network suffered from limitations in storage capacity. The degradation of memory recall with increased storage density is directly related to the presence in the state space of unwanted local minima which serve as basins of flow.
Bachmann, Cooper, Dembo and Zeitouni have studied a system not unlike the Hopfield network; however, they have focused on defining a dynamical system in which the locations of the minima are explicitly known, and for which it is possible to demonstrate that there are no unwanted local minima. Bachmann, C. M., Cooper, L. N., Dembo, A., Zeitouni, O.: "A Relaxation Model for Memory with High Density Storage," Proc. Natl. Acad. Sci. U.S.A., Vol. 84, No. 21, pp. 7529-7531 (1987). In particular, they have chosen a system with a Liapunov function given by ##EQU1## where μ is the N-dimensional state vector and Xj is the jth memory in the N-dimensional state space. This N-dimensional Coulomb energy function defines exactly m basins of attraction to the fixed points located at the charge (memory) sites xj. It can be shown that convergence to the closest distinct memory is guaranteed, independent of the number of stored memories m, for proper choise of N and L. Dembo, A., Zeitouni, O.: "ARO Technical Report," Brown University, Center for Neural Science, Providence, R.I. Bachmann et al. have employed a network implementation of this system which is strictly local however, and does not have some of the desirable characteristics of the distributed Hopfield representations.
The present invention provides a learning algorithm for the N-dimensional Coulomb network which is applicable to single and multi-layer networks and develops distributed representations. The central concept is to define a potential energy of a collection of memory sites. Then each memory site is an attractor of not only test patterns, but other memory sites as well. With proper definition of attractive and repulsive potentials between the memory sites, it is possible to minimize the energy of the collection of memories. With this method, the training time of multi-layer networks is a linear function of the number of layers.
A method for the consolidation of the representation of memories of events that have already been recorded in a layer of cells is first considered. It is shown that this method is applicable to a supervised learning paradigm in which pairs of input and output (or classification) patterns are presented sequentially to the network. It is then shown how the resulting learning procedure develops internal representations incrementally from the layer closest to the input, to the output layer.
For a full understanding of the present invention, reference should now be made to the following detailed description of the preferred embodiments of the invention and to the accompanying drawings.
FIG. 1 is a representational diagram of a neural network illustrating a network implementation of the N-dimensional Coulomb energy. Computation of the memory distances is performed by an "effective" cell and broadcast through non-modifiable synapses, which are shown as white circles in the drawing.
FIG. 2 is a representational diagram of a two-layer neural network for the development of the first mapping D2 =ω1 (D1). Recurrent connectivity is not illustrated.
FIG. 3 is a diagram showing the pattern environment D1 for the XOR problem. Two classes of patterns reside at the corners of the unit square in the activity (or response) space of layer 1.
FIG. 4a is a diagram showing the initial mapping ω1 (D1), plotted in the response space of layer 1, which creates the pattern set D2.
FIG. 4b is a diagram showing the pattern set D2 which consists of only three distinct clusters of patterns in the response space of layer 2.
FIG. 5 is a representational diagram of a three-layer neural network for the complete mapping D3 =ω2 (D2)=ω2 (ω1 (D1)).
FIG. 6, comprised of FIGS. 6a, 6b and 6c, depicts the individual mappings from layer 1 to layer 3 in the neural network of FIG. 5. In this case, each of the two non-linear mappings ω1 (D1) and ω2 (D2) were learned separately. FIGS. 6a, 6b and 6c are diagrams showing, respectively, the layer 1, layer 2 and layer 3 response spaces in the neural network of FIG. 5.
FIG. 7 is a representational diagram of a single layer, N-dimensional Coulomb neural network, illustrating a typical implementation of the learning algorithm according to the present invention.
FIG. 8 is a diagram of a multi-layer neural network of the type shown in FIG. 7, illustrating the connectivity between layers.
Learning In The N-Dimensional Coulomb Network:
Consider a system of m memory sites x1, . . . , Xm in the network response space RN. Let memory sites associated with a particular class, C, be xi (c). Then, employing the energy of Bachmann et al., ##EQU2## in the absence of the test charge μ, the electrostatic potential energy of the configuration of memory sites is: ##EQU3##
First we define an attractive potential energy between memory sites of the same class, and a repulsive potential energy between memory sites of different class. Then we define the Qi such that: ##EQU4## In this case Ψ describes the electrostatic potential energy of a collection of charges of like and different sign, located at the sites x1, . . . , xm in RN.
Then we may relax the constraint that the m memory sites are fixed in RN, by allowing the charges to evolve according to the laws of electrodynamics. Intuitively, the memory sites of same class (but pairwise different charge sign) will be attracted to each other, and memory sites of different class (and same charge sign) will be repelled from each other. We make the definition ##EQU5## where en is the unit vector identifying the nth cell, ωnm is the matrix of afferent synapses to the layer of cells with activity μ, and fmi is the afferent activity copied as memory xi. For simplicity we make a notational change: ##EQU6## Now, motion of the memory sites in the activity space of the layer of cells occurs as a result of changes in the afferent matrix of synapses to the cell layer, ω. We would like to know how to adjust the matrix ω such that the electrostatic energy, Ψ, of the collection of charges (corresponding to the memory sites) is minimized. Thus, in the usual manner we compute the gradient of the potential energy with respect to the weights ωnm : ##EQU7## Thus, for the nth cell in the network, the afferent vector of weights, ωn =[ω1n, . . . ω1K ], must evolve according to: ##EQU8## Relaxation of memory sites according to equation (11) will allow consolidation of memory sites of same class, and separation of memory sites of different class through the minimization of Ψ. That is, the mapping ω evolves in such a way as to collapse the representation of a class in RN while enhancing class distinctions.
Note, however, that with a linear dependence of each cell activity on the inner product of a cell weight vector, a multilayer network will not achieve further consolidation than a single layer. Following the method of Rumelhart et. al. disclosed in Rumelhart, D. E., Hinton, G. E., and Williams, R. J.: "Learning Representations by Back Propagating Errors", Nature, 323, 533-536 (1986), we employ the sigmoid function F to represent cell activity: ##EQU9## Then: ##EQU10## It is easy to show that:
∂/∂ωnm xi =en Fn (1-Fn)fmi,
and thus that: ##EQU11## For brevity we define:
Rij ∂/∂ωnm Rij ≡Δnm (fi,fj).
Then finally, the evolution of the mth synapse of the nth cell follows: ##EQU12## Evolution of the matrix ω, according to equation (13) will minimize the electrostatic potential energy of a collection of m charges by allowing charges of opposite sign (same class) to be attracted to each other and charges of same sign (or opposite class) to be repelled from each other. In other words, this is an alogrithm for the generation of a non-linear mapping from an input space, RN, with m distinct memories of M classes, to a space RN', with n (n<m) distinct memories of M classes. That is, the mapping preserves the class structure of the pattern environment, but consolidates the representation of the environment.
Of interest are possible network implementations of equation (13). The algorithm described relies on the pairwise interactions of the memories. For this reason, each node in the network must "remember" the locations of the memories in its activity space so that the differences Rij may be computed. This form of implementation is not impossible, however it does require the use of more complex nodes than are usually considered in neural networks. An alternative network implementation of this algorithm can be devised which requires nodes which need only remember their activity levels for the previously presented pattern. This method can be implemented through the use of supervised (or unsupervised) learning.
The algorithm described by equation (13) may be easily extended to systems in which the memories do not already exist. That is, we may view this as an algorithm for learning representations from a pattern environment, D. We employ the standard supervised learning procedure in which patterns from D are presented to the network which produces output vectors or classifications of the input. Training consists of adjusting the synaptic state ω to give the correct output classifications.
Equations (13) describes a pairwise interaction between memories; a term in the equation contains factors from patterns i and j. We would like to approximate equation (13) through the accumulation, over time, of the N2 terms. Thus, we must define a system which remembers the response of the layer to the last pattern seen, and is able to then compare it to the response to the next pattern seen. That is, suppose that patterns f(t) and f(t+1) appear on the input lines to the layer of cells. The activity pattern in response to a pattern f(t) is then defined as: ##EQU13## Then, as before, we may define a relaxation procedure on ω, but now as a function of the activity caused by patterns in the environment. For subsequent patterns,
δωnm =(+/-)|x(t)-x(t+1)|-(L+2)Δnm (f(t),f(t+1)) (14)
where we take the negative sign for subsequent patterns of the same class, and the positive sign for patterns of different class.
Equation (14) is a very simple prescription for the acquisition of a representation ω: if two patterns are of the same class, then adjust ω such that their positions in the response space (RN) move closer to each other, if the two patterns are of different class, adjust ω such that the two pattern responses move apart.
To better understand the relationship between equation (14) and the structure of the environment D, we must define D in greater detail. Suppose that the pattern environment consists of m distinct patterns: D=[f1, . . . , fm ]. However, suppose that some patterns occur more frequently than others over time; in particular, suppose that pattern fi occurs with a probability pi. Then if we average equation (14) over the full pattern environment: ##EQU14## That is, the evolution of the system in response to a pattern fi must be computed through an average over the full set of N patterns. This requires the computation of N terms for each pattern. Then evolution of the system for the full environment requires an average of equation (15) over the environment: ##EQU15## Comparison of equations (13) and (16) allows us to identify the charge Qi with the probability pi of a pattern causing a response at the location xi. This identification is intuitively clear: the size of the attractive basin surrounding a memory in the energy defined by equation (1) is a function of the amount of charge at the memory site. Memories of events which are very frequent in D will form larger basins of attraction in RN, and thus be the likely identifiers of new events μ.
We may define an energy according to equation (3) for each layer of a multi-layer network. Then the energy of each layer may be minimized according to (13) (or equivalently equation (16)), independent of the other layers in the system. δω does not depend on an error that is propagated through more than one layer. Convergence to an internal representation progresses from the layer closest to the input, to the layer nearest the output. This system constructs an internal representation layer-by-layer.
Network implementation of equation (16) requires that each cell in the network compare the current activity Fn (fj) with the activity of a previously presented pattern Fn (fi), (implicit in the quantity Δnm (fi,fj)). For pairwise interactions, this requires the computation of N2 terms. Thus, training time for layers nearest to the input will be proportional to the square of the number of patterns in the environment. However, deeper layers in a multi-layer network will train much faster since the number of distinct activity levels will be reduced to the number of distinct classes of events in the environment as one approaches the output layer.
Further, equation (16) requires that each cell have access to the quantity |Rij |. This factor is a measure of the distance between network states for different patterns. This measure is the same for all cells; it is as if a single virtual cell was computing the distance in activity space between subsequent network states and broadcasting this quantity throughout the network.
Interestingly, this mechanism is related to a mean field theory of the visual cortex, as reported in Scofield, C. L.: "A Mean Field Theory of Layer IV of Visual Cortex and its Application to Artificial Neural Networks.: Presented at the IEEE Conference on "Neural Information Processing Systems--Natural and Synthetic", Denver, Colo. (1987). Recent study by Cooper and Scofield of the target layer of visual input to mammalian cortex has found that a mean field approximation of the detailed cortical activity patterns is sufficient to reproduce the experimentally observed effects of deprivation on visual learning. Recurrent cortical connectivity may be replaced by a single "effective" cell which computes the mean cortical activity and broadcasts the signal to all cells of the network. In a similar fashion, the factor |Rij | may be computed by a single effective cell and returned to each cell in the network. This configuration is illustrated in FIG. 1 which shows a two-layer network with a first layer of N cells 10 and a second layer of N cells 12 with an effective cell 14.
To see that this system develops a single internal representation incrementally from one layer to the next, we must look in detail at the representations that develop. We begin with some notation. Let the environment of pattern activity afferent to the ith layer of cells be Di. The matrix of synapses to the ith layer of cells ωi transforms the environment Di to the activity patterns Di+1 :
Di+1 =ωi (Di) (17)
We consider the XOR problem since this has been studied in detail with the generalized delta rule. We begin with a simple two layer network with two cells 16 and 18, and 20 and 22, respectively, in each layer, as illustrated in FIG. 2, in which the input layer is the source of pattern environment Di, which projects to a second layer of two cells through the matrix ω1. (For clarity, mean field connectivity for the computation of |Rij | is not illustrated in FIGS. 2-6.)
The environment consists of the set D1 =[f1 (a), f2 (a), f3 (b), f4 (b)], with f1 (a) and f2 (a) belonging to class "a" and f3 (b) and f4 (b) belonging to class "b": ##EQU16## These four patterns from memory sites in the input space (layer 1) of R2, which reside at the corners of the unit square, are illustrated in FIG. 3.
The locations of these memories in the activity space of the layer-2 cells will depend on the initial values of the matrix ω1. Evolution of ω1 according to equations (13) or (14) will move the memory sites in the activity space of layer-2 until the potential energy of the configuration of memories is minimized. FIG. 4a illustrates the final configuration of activities in layer 2 for the pattern set D1.
Also shown is the location of the two synaptic states ω1 1 and ω2 1 in the input space. Note that with the constraint of a two cell architecture at layer 2, it was not possible for all memories of the same class to settle to a single location in the activity space. Rather, Ψ was minimized through the collapse of two memory sites onto one location and the separation of the other class memory sites from this location.
The activity patterns corresponding to the locations in activity space of layer 2, of the three distinct memories, now form the pattern set D2 for possible third layer cells (FIG. 4b).
FIGS. 4a and 4b show that the memory configuration resulting from the transformation ω1 (D1) is now linearly separable. Memory sites of class "a" may be separated from memory sites of class "b" with a simple linear binary threshold unit. The addition of a single cell 24 as a third layer to the configuration represented in FIG. 2, as illustrated in FIG. 5 (and as represented diagrammatically in FIG. 6) with matrix ω2 produces the transformation D3 =ω2 (D2)=[`a`,`b`,]=[1,0].
Examination of the resulting 3-layer network shows that the two transformations ω1 and ω2 form a single internal representation
ω(D1)=ω2 (D2)=ω2 (ω1 (D1)) (19)
which is identical with that produced by the generalized delta rule. However, here the final transformation was constructed incrementally from ω1 to ω2. This is a significant advantage over the method of Back Propagation since training time will be a simple linear function of the number of layers in the network.
If the average time for convergence at a layer is tc, then the time of convergence of a network with L layers is just Ltc. That is, the time for convergence of a multi-layer network depends linearly on the number of layers in the network.
Discussion of Results:
A learning algorithm for the N-dimensional Coulomb network has been described which is applicable to single and multi-layer networks and develops distributed representations. The method relies on defining attractive and repulsive potentials between the various memories and an electrostatic potential energy of the collection of memory sites. Minimization of this energy through the adjustment of network synaptic states results in a training time of multi-layer networks which is a linear function of the number of layers. Thus, unlike the generalized delta rule, this method appears to scale well with problem size.
This method according to the invention may be applied to the supervised training of networks. With this method the underlying statistics of the pattern environment are explicitly related to the sizes of the basins of attraction. Unsupervised training is a trivial extension in which all patterns are treated as charges of same sign, thus producing mutually repulsive memory sites. Clustering is the natural result of the relative charge magnitudes defined by the statistics of the pattern environment.
The learning algorithm according to the present invention, given by equation (14) may be implemented in an N-dimensional Coulomb neural network in a straightforward manner. Reference is now made to FIG. 7 which depicts a simple, single-layer recurrent network.
The network of FIG. 7 comprises several different types of components which are illustrated with different kinds of symbols. The operation of the network will first be summarized, and will then be described in detail below.
Patterns enter the network via the lines 26 at the top left of FIG. 7. These activity patterns f1 (t), . . . fm (t), . . . fK (t) (real numbers in the range from 0 to 1) enter the network and form modifiable connections ωnm at the small, black, disk intersections 28 at the top of the figure. The large circles 30 numbered 1 . . . n . . . N are the cells of the network which sum activity on their lines of input (stemming from the cell "bodies", upward, in the figure). These cells produce activity levels x1 (t), . . . xn (t), . . . xN (t), respectively (also real numbers in the range from 0 to 1). These activity levels form the input to the "effective" cell 32, shown as a large shaded circle, and exit the circuit along the lines at the bottom right of the figure. The effective cell computes the state space normalized distance |x(t)-x(t+1)|, and broadcasts it to all cells in this layer.
Considering now the various functions of the circuit in detail, we focus initially on the small, black, disk-like connections which represent the connection strength ωnm which are modified in accordance with the learning algorithm of the present invention. Signals f(t) from outside of the circuit arrive via the lines entering at the left. These signals are weighted by the connection strengths ωnm between the line m and the cell n. These weighted signals are then transmitted to the cell bodies where they are summed and transformed into cell activities ##EQU17## where x(t) is a vector comprised of the real number activities x1 (t), . . . xn (t), . . . xN (t), and en is the unit vector.
In addition, each of these connections having connection strengths ωnm must hold the previously seen signal fm (t) for use in the computation of equation (14) set forth above.
The small light squares 34 in the lower right area of FIG. 7 are connections which receive the signals x1 (t) . . . xn (t) . . . xN (t). These signals are stored at these connections and compared to the next successive signals at time (t+1). The square connection, for example the nth connection, then computes the difference (xn (t)-xn (t+1))2.
The effective cell accumulates the N2 difference terms and computes the state space distance |x(t)-x(t+1)|, which is the normalized N-dimensional vector difference, and thereafter a function of the state space difference which is the inverse power: |xn (t)-xn (t+1)|-(L+2), where L is an integer greater than or equal to N-2. Specifically: ##EQU18##
The effective cell transmits the function of the state space distance to all N cells in this layer of the network. These cells each compute the update term to the modifiable connections (equation (14)), and feed this term back to the connections for adjustment. The update term is a product of the quantities |x(t)-x(t+1)|-(L+2)) and Δnm (f(t), f(t+1)), wherein this latter quantity is given by: ##EQU19## where Fn (f(t)) is defined as xn (t).
To compute this quantity, each cell must retain a copy of its last activity level, Fn (t), for comparison to the next activity level, Fn (t+1), so as to compute Fn (f(t))-Fn f(t+1), and the factors: |x(t)-x(t+1)|-(L+2)
In addition, each connection ωnm must retain a copy of the previously seen signal fm (t). These quantities must be available at the connections ωnm so that they may be combined to form the factor: Δnm (f(t), f(t+1)), that is, the cell and its connections must have available the following quantities: ##EQU20##
It will be understood that the circuit of FIG. 7 may be implemented either by analog or digital circuitry. It is perhaps most convenient if each of the elements (connections ωnm, network cells, output connections to the effective cell and the effective cell itself) are each implemented by dedicated digital hardware. The N-dimensional network can also be implemented and simulated by software in a parallel processing computer.
Whereas the preceding discussion focused on the development of a single layer N-dimensional network, once a layer has converged to a representation, additional layers may be added as illustrated in FIG. 8. In fact, such multiple layers may develop concurrently and independently of each other in a multi-layer, N-dimensional Coulomb neural network.
There has thus been shown and described a novel learning algorithm for an N-dimensional Coulomb network which fulfills all the objects and advantages sought therefor. Many changes, modifications, variations and other uses and applications of the subject invention will, however, become apparent to those skilled in the art after considering this specification and the accompanying drawings which disclose the preferred embodiments thereof. All such changes, modifications, variations and other uses and applications which do not depart from the spirit and scope of the invention are deemed to be covered by the invention which is limited only by the claims which follow.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US4752906 *||Dec 16, 1986||Jun 21, 1988||American Telephone & Telegraph Company, At&T Bell Laboratories||Temporal sequences with neural networks|
|US4796199 *||Feb 24, 1987||Jan 3, 1989||Oregon Graduate Center||Neural-model, information-handling architecture and method|
|1||*||A Relaxation Model for Memory with High Storage Density by Charles M. Bachmann et al. Neural Networks, pp. 1 9.|
|2||A Relaxation Model for Memory with High Storage Density by Charles M. Bachmann et al.-Neural Networks, pp. 1-9.|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US4990838 *||Jan 3, 1990||Feb 5, 1991||Atr Auditory And Visual Perception Research Laboratories||Movement trajectory generating method of a dynamical system|
|US5067164 *||Nov 30, 1989||Sep 10, 1991||At&T Bell Laboratories||Hierarchical constrained automatic learning neural network for character recognition|
|US5130563 *||Jun 7, 1991||Jul 14, 1992||Washington Research Foundation||Optoelectronic sensory neural network|
|US5134685 *||Feb 6, 1990||Jul 28, 1992||Westinghouse Electric Corp.||Neural node, a netowrk and a chaotic annealing optimization method for the network|
|US5195171 *||Sep 30, 1991||Mar 16, 1993||Yozan, Inc.||Data processing system|
|US5197008 *||Jan 25, 1991||Mar 23, 1993||Mitsubishi Jidosha Kokyo Kabushiki Kaisha||System for controlling the output power of a motor vehicle|
|US5253328 *||Nov 17, 1989||Oct 12, 1993||Microelectronics & Computer Technology Corp.||Neural-network content-addressable memory|
|US5293459 *||Oct 23, 1991||Mar 8, 1994||U.S. Philips Corporation||Neural integrated circuit comprising learning means|
|US5317741 *||May 10, 1991||May 31, 1994||Siemens Corporate Research, Inc.||Computer method for identifying a misclassified software object in a cluster of internally similar software objects|
|US5371834 *||Aug 28, 1992||Dec 6, 1994||The United States Of America As Represented By The Administrator Of The National Aeronautics And Space Administration||Adaptive neuron model--an architecture for the rapid learning of nonlinear topological transformations|
|US5371835 *||Aug 12, 1992||Dec 6, 1994||Kabushikikaisha Wacom||Inductively coupled neural network|
|US5412670 *||Nov 30, 1992||May 2, 1995||Ricoh Corporation||N-bit parity neural network encoder|
|US5424773 *||Jan 28, 1994||Jun 13, 1995||Kawai Musical Inst. Mfg. Co., Ltd.||Apparatus and method for generating a pseudo camera position image from a plurality of video images from different camera positions using a neural network|
|US5428788 *||Jun 4, 1993||Jun 27, 1995||Siemens Corporate Research, Inc.||Feature ratio method for computing software similarity|
|US5438676 *||Jun 4, 1993||Aug 1, 1995||Siemens Corporate Research, Inc.||Method for adapting a similarity function for identifying misclassified software objects|
|US5440742 *||Apr 6, 1994||Aug 8, 1995||Siemens Corporate Research, Inc.||Two-neighborhood method for computing similarity between two groups of objects|
|US5461698 *||Jun 17, 1993||Oct 24, 1995||Siemens Corporate Research, Inc.||Method for modelling similarity function using neural network|
|US5467427 *||Jun 6, 1994||Nov 14, 1995||Iowa State University Research Foundation||Memory capacity neural network|
|US5467429 *||Jun 28, 1994||Nov 14, 1995||Nippon Telegraph And Telephone Corporation||Neural network circuit|
|US5479574 *||Apr 1, 1993||Dec 26, 1995||Nestor, Inc.||Method and apparatus for adaptive classification|
|US5485621 *||Apr 6, 1994||Jan 16, 1996||Siemens Corporate Research, Inc.||Interactive method of using a group similarity measure for providing a decision on which groups to combine|
|US5493632 *||Aug 27, 1993||Feb 20, 1996||Goldstar Co., Ltd.||Neural network employing a location addressable memory and method for operating the same|
|US5553196 *||Jun 5, 1995||Sep 3, 1996||Yozan, Inc.||Method for processing data using a neural network having a number of layers equal to an abstraction degree of the pattern to be processed|
|US5634063 *||Nov 20, 1995||May 27, 1997||Goldstar Co., Ltd.||Neural network and method for operating the same|
|US5822741 *||Feb 5, 1996||Oct 13, 1998||Lockheed Martin Corporation||Neural network/conceptual clustering fraud detection architecture|
|US5949367 *||Feb 13, 1998||Sep 7, 1999||Alcatel Alsthom Compagnie Generale D'electricite||Device and method for classifying objects in an environmentally adaptive manner|
|US6167390 *||Dec 8, 1993||Dec 26, 2000||3M Innovative Properties Company||Facet classification neural network|
|US6366897||Jul 26, 1999||Apr 2, 2002||Hnc Software, Inc.||Cortronic neural networks with distributed processing|
|US6505168||Aug 16, 1999||Jan 7, 2003||First Usa Bank, Na||System and method for gathering and standardizing customer purchase information for target marketing|
|US6865547||Feb 16, 2000||Mar 8, 2005||Bank One Delaware, N.A.||Customer activated multi-value (CAM) card|
|US7020631||May 21, 2001||Mar 28, 2006||The Chase Manhattan Bank||Method for mortgage and closed end loan portfolio management|
|US7072864||Feb 20, 2003||Jul 4, 2006||Bank One Deleware, N.A.||Customer activated multi-value (CAM) card|
|US7096192||Nov 17, 1999||Aug 22, 2006||Cybersource Corporation||Method and system for detecting fraud in a credit card transaction over a computer network|
|US7212996||Apr 20, 2000||May 1, 2007||Jpmorgan Chase Bank, N.A.||System and method for dynamic, multivariable comparison of financial products|
|US7403922||Nov 2, 2000||Jul 22, 2008||Cybersource Corporation||Method and apparatus for evaluating fraud risk in an electronic commerce transaction|
|US7433840||Mar 14, 2007||Oct 7, 2008||Jpmorgan Chase Bank, N.A.||System and method for dynamic, multivariable comparison of financial products|
|US7467109||May 7, 2007||Dec 16, 2008||Jpmorgan Chase Bank, N.A.||System and method for determining eligibility and enrolling members in various programs|
|US7472088||Jun 1, 2001||Dec 30, 2008||Jpmorgan Chase Bank N.A.||System and method for offering a financial product|
|US7480631||Dec 15, 2004||Jan 20, 2009||Jpmorgan Chase Bank, N.A.||System and method for detecting and processing fraud and credit abuse|
|US7509279||Apr 17, 2006||Mar 24, 2009||Riskmetrics Group, Inc.||Method and system for computing path dependent probabilities of attaining financial goals|
|US7542921||Sep 29, 2000||Jun 2, 2009||Jpmorgan Chase Bank, N.A.||Network-based financial planning system and method|
|US7599879||Mar 21, 2001||Oct 6, 2009||Jpmorgan Chase Bank, National Association||Syndication loan administration and processing system|
|US7606756||Jun 3, 2003||Oct 20, 2009||Jpmorgan Chase Bank, N.A.||Synthetic funds having structured notes|
|US7610231||Nov 1, 2005||Oct 27, 2009||Jpmorgan Chase Bank, N.A.||Synthetic funds having structured notes|
|US7624068||Aug 18, 2004||Nov 24, 2009||Jpmorgan Chase Bank, N.A.||Method and system for dynamically adjusting discount rates for a card transaction|
|US7668777||Jul 23, 2004||Feb 23, 2010||Jp Morgan Chase Bank||System and method for providing instant-decision, financial network-based payment cards|
|US7685064||Apr 19, 2005||Mar 23, 2010||Jp Morgan Chase Bank||Method and apparatus for evaluating a financial transaction|
|US7707192||May 23, 2006||Apr 27, 2010||Jp Morgan Chase Bank, N.A.||Confidence index for assets|
|US7752084||Nov 25, 2009||Jul 6, 2010||Cybersource Corporation||Method and system for detecting fraud in a credit card transaction over the internet|
|US7756896||Apr 7, 2005||Jul 13, 2010||Jp Morgan Chase Bank||System and method for multi-dimensional risk analysis|
|US7774248||Apr 19, 2005||Aug 10, 2010||Jp Morgan Chase Bank||Method and apparatus for managing risk|
|US7831467||Oct 17, 2000||Nov 9, 2010||Jpmorgan Chase Bank, N.A.||Method and system for retaining customer loyalty|
|US7831509||Aug 22, 2005||Nov 9, 2010||Jpmorgan Chase Bank, N.A.||On-line higher education financing system|
|US7844518||Apr 19, 2005||Nov 30, 2010||Jp Morgan Chase Bank||Method and apparatus for managing credit limits|
|US7865427||May 8, 2002||Jan 4, 2011||Cybersource Corporation||Method and apparatus for evaluating fraud risk in an electronic commerce transaction|
|US7890343||Jan 11, 2005||Feb 15, 2011||Jp Morgan Chase Bank||System and method for generating risk management curves|
|US7895098||Mar 1, 2002||Feb 22, 2011||Jpmorgan Chase Bank, N.A.||System and method for measuring and utilizing pooling analytics|
|US7925578||Aug 26, 2005||Apr 12, 2011||Jpmorgan Chase Bank, N.A.||Systems and methods for performing scoring optimization|
|US7925583||Nov 23, 2009||Apr 12, 2011||Jpmorgan Chase Bank, N.A.||Method and system for dynamically adjusting discount rates for a card transaction|
|US7953663||Sep 4, 2003||May 31, 2011||Jpmorgan Chase Bank, N.A.||System and method for financial instrument pre-qualification and offering|
|US7962391||Nov 14, 2008||Jun 14, 2011||Jpmorgan Chase Bank, N.A.||System and method for determining elegibility and enrolling members in various programs|
|US7962396||Feb 3, 2006||Jun 14, 2011||Jpmorgan Chase Bank, N.A.||System and method for managing risk|
|US7970701||Feb 20, 2008||Jun 28, 2011||Cybersource Corporation||Method and apparatus for evaluating fraud risk in an electronic commerce transaction|
|US7974895||Mar 11, 2005||Jul 5, 2011||Jp Morgan Chase Bank||System and method for developing finance rate information|
|US8019678||Nov 3, 2010||Sep 13, 2011||Cybersource Corporation||Method and apparatus for evaluating fraud risk in an electronic commerce transaction|
|US8027914||Jan 5, 2010||Sep 27, 2011||Jp Morgan Chase Bank||System and method for providing instant-decision, financial network-based payment cards|
|US8170952||Aug 3, 2011||May 1, 2012||Jp Morgan Chase Bank||System and method for providing instant-decision, financial network-based payment cards|
|US8175908||Sep 3, 2004||May 8, 2012||Jpmorgan Chase Bank, N.A.||Systems and methods for constructing and utilizing a merchant database derived from customer purchase transactions data|
|US8244629||Jun 24, 2011||Aug 14, 2012||Michael Lewis||Method and apparatus for generating a bi-gram score in fraud risk analysis|
|US8255307||Jan 18, 2011||Aug 28, 2012||Jpmorgan Chase Bank, N.A.||System and method for measuring and utilizing pooling analytics|
|US8285625||Aug 18, 2011||Oct 9, 2012||Jpmorgan Chase Bank, N.A.||Synthetic funds having structured notes|
|US8429006||Jun 18, 2004||Apr 23, 2013||Jpmorgan Chase Bank, N.A.||System and method for offer targeting|
|US8478637||Apr 8, 2008||Jul 2, 2013||Jpmorgan Chase Bank, N.A.||Index for assessing discount potential|
|US8489497||Jan 27, 2006||Jul 16, 2013||Jpmorgan Chase Bank, N.A.||Online interactive and partner-enhanced credit card|
|US8533030||Aug 30, 2004||Sep 10, 2013||Jpmorgan Chase Bank, N.A.||In-bound telemarketing system for processing customer offers|
|US8533031||Sep 17, 2010||Sep 10, 2013||Jpmorgan Chase Bank, N.A.||Method and system for retaining customer loyalty|
|US8577770||Jul 23, 2012||Nov 5, 2013||Jpmorgan Chase, N.A.||System and method for measuring and utilizing pooling analytics|
|US8719078||May 29, 2013||May 6, 2014||Jpmorgan Chase Bank, N.A.||Index for assessing discount potential|
|US8751391||Mar 31, 2003||Jun 10, 2014||Jpmorgan Chase Bank, N.A.||System and process for performing purchase transactions using tokens|
|US8762260||Aug 14, 2012||Jun 24, 2014||Jpmorgan Chase Bank, N.A.||Systems and methods for performing scoring optimization|
|US8782076||Oct 19, 2010||Jul 15, 2014||Jpmorgan Chase Bank, N.A.||System and method for gathering and standardizing customer purchase information for target marketing|
|US20010029477 *||May 21, 2001||Oct 11, 2001||Chase Manhattan Bank||Method for mortgage and closed end loan portfolio management|
|US20010054022 *||Mar 21, 2001||Dec 20, 2001||Louie Edmund H.||Syndication loan administration and processing system|
|US20020052776 *||Nov 19, 2001||May 2, 2002||First Usa Bank, Na||System and method for gathering and standardizing customer purchase information for target marketing|
|US20020147669 *||Jun 1, 2001||Oct 10, 2002||Taylor James W.||System and method for offering a financial product|
|US20040148241 *||Jun 23, 2003||Jul 29, 2004||Qi Thomas J.||Method of evaluating a portfolio of leased items|
|US20040148566 *||May 13, 2003||Jul 29, 2004||Jp Morgan Chase Bank||Method to evaluate project viability|
|US20040236647 *||May 23, 2003||Nov 25, 2004||Ravi Acharya||Electronic checkbook register|
|US20050060252 *||Oct 30, 2003||Mar 17, 2005||Andrew Doddington||Graphical software tool for modeling financial products|
|US20050182713 *||Sep 30, 2004||Aug 18, 2005||Giancarlo Marchesi||Methods and systems for the auto reconsideration of credit card applications|
|US20050288954 *||Jun 6, 2005||Dec 29, 2005||Mccarthy John||Method, system and personalized web content manager responsive to browser viewers' psychological preferences, behavioral responses and physiological stress indicators|
|WO1992013306A1 *||Dec 30, 1991||Aug 6, 1992||Univ Boston||Vector associative map system|
|U.S. Classification||706/25, 706/32, 706/31, 706/27|
|International Classification||G06F15/18, G06N3/08, G06N3/04, G06N99/00|
|Cooperative Classification||G06N3/08, G06N3/04, G06K9/6274|
|European Classification||G06K9/62C1D1R, G06N3/08, G06N3/04|
|Jan 19, 1988||AS||Assignment|
Owner name: NESTOR, INC., A CORP. OF DE, RHODE ISLAND
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:SCOFIELD, CHRISTOPHER L.;REEL/FRAME:004836/0109
Effective date: 19880113
|Jul 23, 1993||FPAY||Fee payment|
Year of fee payment: 4
|Sep 2, 1997||SULP||Surcharge for late payment|
|Sep 2, 1997||FPAY||Fee payment|
Year of fee payment: 8
|Sep 9, 1997||REMI||Maintenance fee reminder mailed|
|Jul 3, 2001||FPAY||Fee payment|
Year of fee payment: 12
|Sep 18, 2006||AS||Assignment|
Owner name: U.S. BANK NATIONAL ASSOCIATION, CONNECTICUT
Free format text: SECURITY AGREEMENT;ASSIGNOR:NESTOR, INC.;REEL/FRAME:018260/0594
Effective date: 20060525